..\
Cognition, 9 (1981) 197-236
@Elsevier Sequoia S.A., Lausanne - Printed in The Netherlands
1
Grammatical description versus configurational arrangement
in language acquisition: The case of relative
clauses in Japanese*
KENJI HAKUTA"*
*
Yale University
A bs trac t
The description o f sentence complexity in terms o f grammatical relations or
in terms o f configurational properties o f their surface structure is a basic
issue in developmental psycholinguistics. The problem has been investigated
in English, but these studies provide little insight because o f the peculiar
properties of word order in the language. A series o f experiments with Japanese children is reported in which the configurational demands of the sentence
are shown to be the critical factor.
TI
I
I
1
Sentences can be described with respect to the grammatical roles of their
constituents, or in terms of the linear arrangement of the words. The extent
to which the parameters specified by these alternative conceptions constrain
children's language is important in determining the nature of the preparedness with which humans begin the task of language acquisition. Although the
question arises in all aspects of investigation in syntactic development, it has
perhaps been most explicitly asked in the study of relative clauses.
In this paper, I will first provide a general characterization of relative
clause structures. This will be followed by a brief review of studies on relative
clause comprehension with children in English, where it will be shown that
the evidence is inconclusive on both empirical and theoretical grounds. The
argument between views emphasizing grammatical description versus linear
*This research was supported by the Peter B. Livingston Fellowship Fund of the Harvard Medical
School, NSF Grants BNS73-09150 and GSOC73-09150 to Dr. Roger Brown, and a dissertation fund
from the Department of Psychology and Social Relations, Harvard University. I would like t o thank
Roger Brown, Jill and Peter de V i e r s , Haj Ross, Helen Tager-Flusberg, Steven Pinker, and two anonymous reviewers for helpful comments and discussion. Even more, I would like to thank my mother,
Emiko Hakuta, for assistance in collecting the data and for general wisdom. Finally;thanks are due to
MI. Yokoyama, Ms. Asami, Ms. Toi, Ms. Abe, and other staff of the day care center where this research
was conducted.
**Requests for reprints should be sent to Kenji Hakuta, Department of Psychology, Yale University,
Box 11A Yale Station, New Haven, CT 06520, U.S.A.
198
Kenji Hakuta
arrangement is best resolved in languages where word order is flexible, one
such language being Japanese. Thus, a series of experiments with Japanese
children will be reported in which the two variables are independently manipulated. Across all experiments, in both comprehension and production, the
surface configurational properties of sentences are shown to be the critical
parameter.
Logical possibilities of sentences with relative clauses
Relative clauses consist of a head' noun and a relativized sentence. The head
noun has a co-referent within the relativized sentence from which, linguistically speaking, it has been extracted. The head noun, placed external t o the
sentence, can come in two positions, either before or after the relativized
sentence, sketched as follows:
(a)
(b)
HEAD NOUN [ RELATIVIZED SENTENCE]
[RELATIVIZED SENTENCE] HEAD NOUN
Since these complex noun phrases are embedded within sentences, a description of how they interact with the matrix sentence is also called for.
Greenberg (1963) has noted that an overwhelming majority of the world's
languages place the subject before the object in the basic underlying order.
This is a strictly statistical generalization, with exceptions such as Tagalog
(Schachter, 1976), but for present purposes, we will assume it be a universal
fact. This leaves us with three logically possible basic orders: Subject-VerbObject (SVO), Subject-Object-Verb (SOV), and Verb-Subject-Object
(VSO). These are in fact very common basic orders among the world's langu ages.
Since the matrix role of the complex noun phrase can be either Subject or
Object (Subject Matrix and Object Matrix), the interaction of the language
type (SVO/SOV/VSO) with the matrix role (Subject Matrix/Object Matrix)
and head noun position (HN[RS] /[RS] HN) creates a variety of different
sentence configurations, appearing in Table I. The sentence configurations
in which the [ R S ] element is flanked on both sides by other constituents
are the centerembedded sentences, while those with [RS] at the beginning
or the end of the string are referred to as left-branching or right-branching.
As it turns out, the world's languages are not evenly distributed in the grid.
The majority of SOV languages have the [ R S ]HN head noun position, while
VSO languages have the HN[RS] position (Greenberg, 1963; Lehmann,
1973). Kuno (1 974) gives an account of this universal, derived from the fact
.
__
-
Relative clauses
Table 1 .
Table 2.
199
Logical possibilities of sentence configurations as a function of the basic
order o f the language, the matrix role of the complex noun phrase, and the
position o f the head noun
Basic order
Matrix role
Head noun position
svo
Subject Matrix
Object Matrix
tJ [RS] -V-N
N-V-N [RS]
[RS] N-V-N
N-V-[RS] N
sov
Subject Matrix
Object Matrix
[RS] -N-V
N-N[RS] -V
[ RS] N-N-V
N-[RS] N-V
vso
Subject Matrix
Object Matrix
V-N[ RS] -N
V-N-N[ RS]
V-[RS]N-N
V-N-[RS] tJ
Logical possibilities of complex noun phrase configurations as a function of
the basic order of the languuge and the position 0.f the head noun
Basic order
Focus
Head noun position
svo
sov
VSO
that centerembedding reduces the comprehensibility of sentences (e.g.,
Miller, 1962; N. Chomsky, 1961). Thus, the tendency for languages is to
seek the head noun position that would result in a lower probability of ending up with center-embedded structures. Antinucci et al. (1979) also provide
a perceptually-based explanation for facts about the diachronic change in
relative clause structure.
In addition to a consideration of the relative clause position in the matrix
sentence, however, a thorough investigation must explore the implication of
the string of elements within the relativized clause and its interaction with
matrix role. Assuming that the basic order of the sentence is preserved in the
relative clause, SVO, SOV and VSO will yield different configurations
depending on (1) whether the subject or the object noun phrase is relativized,
and (2) whether the head noun comes before or after the relativized clause.
The logical possibilities resulting from this combination are outlined in Table
200
Kenji Hakuta
Table 3.
Configurations of nouns and verbs as a function of the basic order of the language, the position of the head noun, the matrk role and head noun focus
Basic
order
Sentence
type
svo
ss
so
os
00
sov
ss
so
os
00
VSO
ss
so
os
00
Head noun position
N [RSI
[RSI N
N[VN)-V-N
N [ N V )-V-N
N-V-Ij[ V N ]
N-V-N[NV]
[ VN] Ij-V-N
[NV] N-V-N
N-V-[ VN]bj
N-V-[NV]
N [ N V ]-N-V
N[NV] -N-V
N-N[NV] -V
N-N[NV] -V
[ N V ]N_-N-V
[NV] N-N-V
N- [NV] Ij-V
N- [ N V ]N-V
V - e [ VN] -N
V-N[ V N ]-N
V-N-N[ VN]
V-N-N[ VN]
'
V-[ VN]N-N
V-[ VN]N-N
V-N-[ VN]N
V-N-[ VN]
2. There are two noteworthy interactions in Table 2. First, the order of the
noun and verb within the relative clause depends on the interaction between
the basic order of the language and the focus. Thus, for SVO languages, the
order for subject focus is [ V N ] , but it is [NVI for object focus. On the
other hand, for SOV and VSO languages, the order remains constant: [ N V ]
for SOV languages and [ V N ] for VSO languages. Second, whether the order
of subject preceding the object is preserved or not depends on the interaction between focus and the position of the head noun. For all three basic
language orders, the subject-before-object order is preserved in subject focus
when the head noun is to the left of the relative clause. The order is preserved
in object focus when the head noun is to the right of the relative clause.
Now, putting together the two logical possibilities for embeddedness as a
function of the matrix role and focus, we end up with Table 3. The point of
the table is that a variety of sentence configurations are possible for any
given sentence type depending on the basic order of the language and the
position of the head noun. It serves to place in context the present line of
research to be described, which was intended to help clarify issues raised by
research conducted on the four sentence configurations in the uppermost
left comer of the table, representing English. The following section describes
the logic underlying the studies of relative clause comprehension in Englishspeaking children, and the results obtained.
Relative clauses
201
Relative clauses in English-speaking children
English is an SVO language, with the head noun of relative clauses on the left.
The following are examples of the four logically possible sentence types in
English (where the first letter refers to the grammatical role of the complex
noun within the matrix sentence, and the second letter refers to the role of
the head noun within the relative clause):
(SS) The duck [that licked the frog] bit the pig.
(SO) The duck [that the frog licked] bit the pig.
(OS) The duck licked the frog [that bit the p i g ] .
(00) The duck licked the frog [that the pig b i t ] .
These four sentence types have been the center of some controversy recently
over what the variables are that determine their psychological complexity for
children.
The controversy exists at both the theoretical and empirical levels. Theoretically, the explanations that have been proposed can be divided into two
distinct classes depending on their assumptions. One class bases itself on the
assumption that the grammatical description of the sentences is the important variable determining their relative complexity. This view takes the grammatical role of the variables involved in these sentences, namely the matrix
role and the focus, to be the determining factor. The surface representation
of these sentences is seen as not relevant. The second class of explanations
homes in exactly on the factors considered irrelevant by the grammatical
description viewpoint, namely the configurational properties of constituents
in the surface structure of sentences. Under this view, the grammatical description of a sentence is seen as a useful summary of the ingredients that go
into yielding configurational features of the sentence. While theories based
on grammatical descriptions would make predictions for cross-linguistic
validation that are universal for the four sentence types, theories based on
the configurational properties of sentences would make predictions that are
based on an interaction between these grammatical descriptions and the
typological characteristics of particular languages as outlined in Table 3,
since this interaction determines the configurational properties of the sentences. On the empirical level as well, there has been an alarming discrepancy
in the results obtained across the several studies in English. I will describe
below the nature of both the theoretical and the empirical controversy and
show how the issue can be resolved, naturally, through an investigation of
Japanese .
Sheldon (1974), who is the major proponent of the view emphasizing
grammatical description as the variable underlying complexity, has argued
202
Kenji Hakuta
that the primary determinant of complexity for these sentences is whether
the grammatical functions of the two variables that defme them are the same
or different. Under Sheldon’s Parallel Function Hypothesis, sentences are
easier to comprehend when the grammatical function of the head noun within the relative clause and the grammatical function of the complex noun
phrase within the matrix sentence are the same than when they are different.
Thus, S S and 00 sentences should be easier than OS and SO. Apparently,
the theoretical motivation underlying Parallel Function is that it constitutes
a significant linguistic generalization not just for relative clauses, but also for
pronominal co-reference and coordination reduction (Sheldon, 1974).
Another reasonable prediction that assumes the importance of the grammatical description of the sentences is based on Keenan and Comrie’s (1 972)
accessibility hierarchy of noun phrases from relative clauses. After surveying
over 40 languages, the authors report a hierarchy in which languages allow
relativization of noun phrases. The hierarchy is: Subject, Direct Object, Indirect Object, Object of Preposition, Possessive Noun Phrase, and Object of
Comparative Particle. Reading this hierarchy from left to right, any language
that allows accessibility of a given noun phrase in the hierarchy will also
allow accessibility of all noun phrases to the left of it. Thus, if a language
allows an Object of Preposition to be relativized, then it will also allow Subject, Direct Object, and Indirect Object. Some languages, such as Tagalog,
only allow the Subject. This putative universal assumes an important role for
developmental psycholinguistics if, as some have speculated, universal rules
should appear earliest in children’s language (e.g., Ross, 1973; McNeill,
1966). Keenan (1975) has expressed his own interest in the relevance of his
hierarchy for adult language performance by investigating the relative frequencies of the various members of the Accessibility Hierarchy in written
English. It would be a natural extension of Keenan’s investigation t o see if
children will fmd relative clauses more difficult as they move down the hierarchy. With reference to the four sentence types mentioned earlier, then, the
Accessibility Hypothesis would predict the Subject Focus sentences ( S S , OS)
to be easier than the Object Focus sentences (SO, 00).
There are several competing explanations for the relative complexity of
these sentences that take into account their configurational properties to
varying extents. One account (Slobin, 1973), consonant with data from
spontaneous speech production and imitation of relative clauses by children,
considers sentences with interruption of the main clause to be more difficult
than those without interruption. Thus, the center-embedded SS and SO
should be more difficult than OS and 00, which are right-branching. The
difficulty with center-embedded sentences is presumably predicted owing t o
the heavy load placed on memory by the interruption of the main clause. As
~
-~
- ...
_
. - .-
-
_
_
Relative clauses
203
.
Sheldon (1 974) points out, however, most studies on center-embedded sentences with adults concerned multiple center-embedded structures, which are
considerably more difficult than single center-embeddings. However, it is
possible that for children, even a single center-embedding would cause difficulty. This will be called the Embeddedness Hypothesis.
Smith (1 974) proposes an explanation that is yet closer to the configurational properties of each sentence. His hypothesis is based on whether the
sentences correspond to Bever’s (1970) Noun-Verb-Noun (NVN) strategy,
in conjunction with the Minimal Distance Principle (MDP) (C. Chomsky,
1969). The MDP was originally proposed by Chomsky to account for
children’s comprehension of sentences with complementizers, but can be extended to relative clauses where constituents are “missing”. The MDP claims
that when children find a missing noun phrase, they will assign the most
recent noun phrase to its location. Smith’s NVN/MDP Hypothesis predicts
that OS should be easiest since NVN applies to the initial segment, and the
MDP assigns the second noun of that sequence as subject of the remaining
VN, yielding the correct interpretation. On the other hand, NVN correctly
interprets the initial segment of S S , but MDP incorrectly assigns the second
noun as the subject of the remaining VN. The NVN strategy correctly interprets the initial segment of 00 sentences as well, but MDP does not apply.
Since MDP mis-applies for SS and fails t o apply to 00, it is not possible to
predict differences between the two, but both should be more difficult
than OS. And finally, the most difficult should be SO, since neither NVN
nor MDP can apply.
Tavakolian (1 978) proposes what I see as an even more local explanation
based on the configurational properties of the sentences. She predicts that
children will interpret these sentences as if they were conjoined. Her predictions work well for SS and OS, both of which have the configuration
NVNVN. These configurations’ could be interpreted by the child as having
the conjoined structure NVN-and-VN. For S S , the sentence The iguana that
bit the tortoise kissed the pigeon, would be perceived by the child as The
iguana bit the tortoise and kissed the pigeon. For S S , the sentence The squirrel hit the lizard that kissed the snail would be perceived as The squirrel hit
the lizard and kissed the snail. A conjoined interpretation would yield the
correct response for SS but would cause an erroneous interpretation of OS.
Tavakolian does not make explicit the implications of conjoined clause analysis for the other two configurations SO and 00, her only claim being that
the conjoined clause analysis would not apply well to these sentences.
The data relevant to the predictions outlined above can be found in studies
by H.D. Brown (1 97 l), Sheldon (1 974), Tavakolian (1 978) and de Villiers
et al. (1979). With the exception of Brown’s study, the studies used an
.
..
204
Kenji Hakuta
act-out comprehension procedure using toy animals. Brown used a picturecued comprehension paradigm, where the child was asked to choose one of
two possible pictures depicting the appropriate action. The results of each
study are displayed in Fig. 1, with Subject and Object Focus sentences plotted separately, and the Matrix Role on the abscissa. The dependent measure
is the percent of total score. For the act-out comprehension studies, since a
breakdown of the response types was available in the published tables, the
percentage of correct acting out of individual clauses out of the total number
of clauses (two per sentence) was calculated for each sentence type. For
H. D. Brown’s (1971) study, the measure is what he reports, namely the
percentage of correct picture choice, for which chance level is 50%.
Figure 1.
Summary of comprehension data from four English studies on SS, SO, OS
and 00. Closed circles represent subject focus, open circles represent object
focus. Sentence type is determined in conjunction with matrix role represented on abscissa. Measure for H D. Brown’s study is per cent of correct choice
of picture, where chance is 50%. For other studies, measure is per cent of
individual clauses acted out correctly. Adapted from H. D. Brown (1971),
Sheldon ( I 9 74), Tavakolian ( I 9 78)and de Villierse t a1 ( I 9 79).
n
S
0
S
0
MATRIX
S
ROLE
0
Relative chuses
205
.
Although at first glance the results across the studies appear consistent, a
close examination reveals that the only common finding is that SO is poorly
comprehended. This result can be construed as evidence for all the hypotheses outlined above: it is not parallel function, it is object rather than subject focus, it is centerembedded, it is not amenable to NVN and MDP, and
it cannot be interpreted as a conjoined structure! Leaving SO aside, the following observations can be made:
1.
2.
3.
S S is easier than OS in Brown (1 972), Sheldon ( 1 974) and Tavakolian
(1 978), but they are equivalent in de Villiers et al. (1 979).
00 is easier than OS in Sheldon, OS is easier than 00 in Brown, and
they are about equivalent in Tavakolian and in de Villiers et al.
SS is easier than 00 in Brown and in Tavakolian, but they are equivalent in de Villiers e t al. and in Sheldon.
The fact that such inconsistencies exist across these studies is quite disturbing, to say the least. Sheldon, Tavakolian and de Villiers et al. all used
identical methodology and highly similar lexical items, in fact all used animals. The subject populations appeared comparable.
The Parallel Function Hypothesis finds strong support in Sheldon’s own
data, since she finds equally good comprehension of SS and 00, and poor
comprehension of OS and SO. The data of Tavakolian and of de Villiers et
aZ., however, are not particularly supportive. In both these studies, there is
roughly equal performance on OS as on 00. In addition, in Tavakolian’s
data, the discrepancy between S S and 00 is very large, which Parallel Function cannot explain. H. D. Brown’s data are quite contradictory to Parallel
Function’s predictions, where OS i s better comprehended than 00.
The Embeddedness Hypothesis is embarassed by the fact that in all but
the de Villiers et al. data, SS is better than OS. This also embarasses the
NVN/MDP Hypothesis, which makes the same prediction. However, the
latter’s redeeming feature is that it explicitly predicts poor performance on
SO. This prediction is supported by all studies.
The Conjoined Clause Hypothesis predicts successfully that S S would be
better comprehended than OS, supported in Brown, Sheldon and Tavakolian.
It does not make any predictions about the relative difficulties of SO and
00, however.
Given such a conflicting array of data, surely no single hypothesis will account for all the data. However, a reasonable account might be formulated
from the Conjoined Clause Hypothesis, which successfully accounts for SS
being superior to OS, and the NVN/MDP Hypothesis, which accounts for SO
being the most poorly comprehended. The Conjoined Clause Hypothesis was
..
- - ..
206
Kenji Hakuta
originally formulated by H. D. Brown (1971). He wrote, “the fact that [SSI
was apparently easier than [OS] may be explained by the fact that in [ SSI
the first subject can also correctly act as the subject of the second verb, as in
a sentence with a single subject and a compound verb, making it easier for
the child to perceive the relationship expressed” (pp. 193 1-2). Tavakolian
(1978) saw the implications of this in accounting not just for good performance on SS but also for the errors in interpreting OS. The Conjoined Clause
Hypothesis accounts well for the superiority of SS over OS. It might be described in more general terms as the child automatically assimilating a given
input sentence into a structure that s/he already knows. It so happens that
the conjoined clause is a good interpretation for a string with the structure
NVNVN, for Enghsh, at least.
When it comes t o 00 and SO, however, the conjoined clause is not a good
match with the sentence configurations, and there is no reason to expect the
child to force such an interpretation on these configurations. In the case of
00, the sequence NVNNV can be interpreted as an initial NVN segment
with an additional NV remaining. We would expect the child to perform
quite well on this sentence because the initial NVN sequence is straightforward. For the remaining NV sequence, given that the child correctly interprets the fragment as agent-action, the chance of finding the correct patient
is 50%. This is the same argument as would be made for SS and OS, except
that in those cases, the missing noun phrase is the subject of a sequence VN.
If there were no conjoined clause interpretation, 00 should be on an equal
footing with SS and OS. However, the conjoined clause interpretation pushes SS upwards and OS downwards on the scale of probability for correct
responses, leaving 00 in between. And fmally, for SO, no existent interpretation is readily available to the child, and so we do not expect good performance.
This explanation, which we will call the Configuration Hypothesis, is a
more general hypothesis of which Tavakolian’s Conjoined Clause Hypothesis
and Smith’s NVN/MDP Hypothesis might be considered special cases. The
Configuration Hypothesis essentially predicts the same results as Parallel
Function, except that it distinguishes between SS and 00 and between OS
and SO, while Parallel Function does not. But even if the two hypotheses
were not confounded in English, it would not be terribly meaningful to pursue the validity of these hypotheses with further studies until some account
of why the data across the various studies are so inconsistent with each other.
It is entirely possible that, had a double-blind procedure been used, we
would not be worrying over the problem presently.
The Parallel Function falls under the general category of explanations that
looks towards the grammatical description of the sentences as the variables
-~
~
-
-
~
~
_
_
..
Relative clauses
207
predicting sentence complexity. The Configuration Hypothesis, on the other
hand, seeks to explain sentence complexity in terms of the configurational
properties of the sentence, which is a function of the parameters of language
described at the beginning of this section. Refemng back to Table 3, it is
possible to see how, by looking at other languages, the two hypotheses can
be unconfounded. I am proposing to determine the correct account for
English by testing the assumptions underlying the hypotheses in other languages. The study of both SOV and VSO languages is important. Taking a
VSO language with the head noun to the left of the relative clause (bottommost left comer of the table), OS and 00 are right-branching while S S and
SO are center-embedded. Now, taking an SOV language with head noun on
the right of the relative clause, S S and SO are now left-branching while OS
and 00 are center-embedded. Notice that the sentence type that ends up
with the center-embedded configuration shifts depending on the basic order
of the language. For VSO, S S and SO are center-embedded, while for SOV,
OS and 00 are center-embedded. Thus, if grammatical description of sentences were important in predicting sentence complexity, then crosslinguistically, sentence type should have an effect regardless of the particular
configuration that the given type takes in a given language. On the other
hand, if configuration were important, one expects variation of complexity
for a given sentence type across languages depending on the configuration
that it takes, such as center-embeddedness.
It turns out that English is a particularly bad language from the viewpoint
of teasing apart the appropriateness of grammatical description and sentence
configuration. There is an inherent confounding of these two descriptions
because of the rigid word order required by English to signal the grammatical
functions of constituents. Given a grammatical description, the configuration
of the sentence is determined. The present study unconfounds these explanations by looking at an SOV language, Japanese.
Relative clauses in Japanese: experiment one
Unlike English, where word order determines the grammatical role of nouns
in a sentence, Japanese signals grammatical role through postpositional particles. Thus, although the predominant, canonical order of a sentence is
Subject-Object-Verb (SOV), that order is free in the sense that word order
is not essential for assigning grammatical role. The major constraint on word
order in Japanese is that the main verb must be sentence-final (Kuno, 1973),
but even this constraint can be violated through dislocation of a constituent
to the right of the verb.
208
Kenji Hakuta
Table 4.
Complex Japanese sentences with relative
clauses in both SOV and OSV orders
Sentence
type
Order
sov
osv
~~
ss
so
os
00
[N-o V ]N-ga N-o V
[N-go VI E g a N-o V
Nga [N-o-V ]N-o V
N-ga [N-go V]’&o V
N-o [N-o V ]N-ga V
N-o [N-ga VI bj-ga V
[N-o V ] N-o N-ga V
[N-go V ]N-o N-ga V
Linguists working in Japanese have not explicitly worked out the full implications for meaning of the word order change, but in general, it is considered an optional rule analogous to extraposition in English. N. McCawley
(1976) formulates the word order change in terms of Ross’s (1 967) “scrambling rule” for Latin, which simply interchanges the position of noun phrases
under the condition that it is postcyclic. One important function of the
scrambling rule in Japanese, as N. McCawley (1976) points out, is to make
multiple center-embedded sentences comprehensible.
Since Japanese is a language with flexible word order, it allows for more
than one configuration for each sentence with a given grammatical description. Thus, taking the middle row of the right column in Table 3 representing Japanese, each of the four sentence types which are represented in the
SOV order there can take the OSV order, which yields a different configuration. Thus, within a language like Japanese, it is possible to unconfound
grammatical description from sentence configuration. It is ideal when these
factors can be separated out within a language rather than across languages,
since languages can differ in so many ways other than the critical variable
one is attempting t o isolate.
The possible sentences in Japanese are listed in Table 4.The configurations on the left column are the same as those representing SOV/[Ris] N in
Table 3, with particles added. On the right column are the configurations
that result when the sentences on the left column are in the OSV order. It
should be pointed out that the only syntactic cue indicating relativization in
Japanese is the constituent order, since there are no relative pronouns in
Japanese. There is generally a slight pause following the head noun, although
I am not aware of any systematic attempts to outline such prosodic cues in
Japanese. The grammatical functions of the different nouns in a sentence
containing a complex noun phrase are signalled, as in simple sentences,
through particles. The particle on the head noun of the relative clause signals
the role of the complex noun phrase within the matrix sentence. Thus, the
__
-
.
.
_ _
Relative clauses
209
role of the head noun within the relative clause is signalled solely by the
particle on the noun phrase within the relative clause. If that noun is marked
by -gaythen the head noun is the object. If it is marked by -9,then the head
noun is the subject.
The sentences in Table 4 have interesting properties that help unconfound
the variables of grammatical description and sentence configuration. If grammatical description were important, one would not expect there to be any
variation in performance as a function of the different configurations that
each sentence type takes. On the other hand, if configurational factors were
important, then we would expect differential performance for each sentence
type with respect to the configuration. Specifically, OS/SOV, OO/SOV, S S /
OSV and SO/OSV are centerembedded, and they all have a sentence-initial
NNV sequence which mimics the simple sentence and could lead to an erroneous interpretation of the sentence. The "V sequence for the respective
sentences take the following forms:
os/sov
oo/sov
ss/osv
so/osv
N-ga N-o V . . .
N-ga N-ga V . , .
N-o N-o V . . .
N-o N-ga V ...
As can be readily seen, OS/SOV mimics the SOV simple active sentence,
while SO/OSV mimics the OSV simple active sentence. Previous experiments
with simple sentences (Hakuta, 1977) showed that children found the SOV
order easier to comprehend than OSV. Thus, the prediction was that the
tendency to choose the fust noun as the agent would be strongest for OS/
SOV and weakest for SO/OSV. In addition, for all the sentences, it was predicted that the first two nouns would be erroneously interpreted as involved
in a relationship, and that the trend would be towards interpreting the first
noun as the agent in all the sentences.
The left-branching sentences, SS/SOV, SO/SOV, OS/OSV and OO/OSV
do not present possibilities for such erroneous interpretations. Thus, a prediction based on considerations of configurational properties'of the sentences
predicts superior performance on the left-branching over the center-embedded sentences. Sheldon's Parallel Function Hypothesis would predict better
performance on S S and 00 over OS and SO, independent of whether they
are in the SOV or OSV order. The hypothesis could tolerate a main effect
for word order, but it would be embarassing to the hypothesis if there were
an interaction between sentence type and the word order of the sentence.
More generally, such an interaction would be problematic for any theory of
complexity based on the grammatical description of the sentences.
210
Kenji Hakuta
In addition to the above predictions, it was hypothesized that there would
still be a main effect for word order. SOV should overall be better comprehended than OSV, as was found for the simple sentences.
Finally, the Accessibility Hierarchy Hypothesis can be tested by assessing
whether Subject Focus sentences are better comprehended than Object Focus
sentences. It provides a particularly critical test of the hypothesis since in
Japanese, Object Focus takes the [ SV] 0 order while the Subject Focus takes
the [OV] S order.
Subjects
Subjects were 12 children between the ages 5;3 and 6;2. In all experiments
reported in this paper, they were in a public day care center in Tokyo, Japan.
Materials and procedure
Sixteen sentences were constructed from factorial combinations of two
levels of Matrix Role (Subject/Object), two levels of Focus (Subject/Object),
two levels of Word Order (SOV/OSV), and two replications. The nouns were
all animals (e.g., alligator, gorilla, camel, panda) randomly assigned to sentence frames. Verbs were ones which required distinct actions (e.g., kicked
[kettal , hit [butta] , licked [nameta] ).
For each sentence, the child was asked to act out the actions on a stage on
which the animals were placed. The experimenter read the sentences at a
slow, natural speed with short pauses following both the subject and the object of the matrix sentence. The experimenter, my mother, was blind t o the
hypotheses of the study, to the extent that mothers can be expected to remain blind to the devious schemes of their children.
Results and discussion
Overall analysis was conducted by a 4-way ANOVA with Subjects crossed
with the repeated measures factor, Sentence, nested within a cross of Matrix
Role (Subject/Object) by Focus (Subject/Object) by Word Order (SOV/
OSV)' . None of the main effects proved significant (for Word Order, F' <
1; for Matrix Role, F ' [ 1, 21 = 1.960, p > 0.10; for Focus, F ' [ 1, 81 = 1.744,
' Following Clark's (1973) suggestion, both subjects and sentences are treated as random effects,
using procedures outlined in Winer (1971). Quasi-F-ratios(F') are reported for ANOVAs, and individual comparisons are made through r'. Whenever means are reported, there are two standard deviations.
S, refers to the variability across subjects, while S, refers to variability across sentences.
___
- .
- .
. ..
.
.
..
Relative clauses
Figure 2 .
2 11
Significant Matrix Role by Word Order interaction.
1.3
W
g
gy
1.2
1.1
a
a
r
w
-9
8
.8
f
.7
a
object matrlx
1.0
subject matrlx
v)
Y
~~~~
~
osv
sov
ORDER
p > 0.10). There was a highly significant Matrix Role by Word Order interaction ( F ‘ [1, 51 = 3 1.492, p < 0.005). The 3-way interaction of Matrix Role
by Focus by Word Order ( F ’ [1,21 = 3.109) approached the alpha level of
0.10, for which the critical value is 3.78.
The significant Matrix Role by Word Order interaction is shown in Figure
2. Since sentences with Matrix Role Subject are left-branching when in the
SOV order but centerembedded when in the OSV order, and the reverse is
the case for Matrix Role Object sentences, the interaction indicates that
Japanese, children find centerembedded sentences more difficult to process
than left-branching sentences. This result is a rather embarassing fact for any
theory which subscribes solely to the grammatical description of sentences
.as predictors of sentence complexity. Thus, the Parallel Function Hypothesis,
without major modification in which sentence configuration is incorporated,
is untenable.
Analysis of individual responses revealed that poor perfoimance on the
center-embedded sentences was mostly due to the erroneous interpretation
of the initial NNV sequence as a simple sentence, as was predicted. Table 5
shows the breakdown of the frequency of responses involving the first two
nouns for each of the four sentence forms. When the response involved the
first noun being the agent of the action, it was coded as “1 -2”, and appears
as such in the Table. When the second noun was the agent, it was coded as
“2-1”. Since there were two responses for each sentence form per child,
each cell has a possible total of 24. Looking at the column totals, it is evident
that most responses across the sentence forms involved interpreting the NNV
21 2
Kenji Hakuta
Table 5 .
Frequencies of responses involving the first two nouns in the centerembedded
sentences
N-gaN-o V...
N-gaN-ga V ...
OSlSOV
OOlSOV
ss10sv
1-2
2-1
23
1
14
8
15
I
14
6
Total
24
22
22
20
Response
N-oN-o V...
N-oN-ga V...
so/osv
“1-2” refers to responses where rust noun is the agent, “2-1” to where second noun
is the agent.
Table 6.
Distribution of responses on the subordinate clause in left-branching sentences
Response
Subject focus
Object focus
SSlSOV
os/osv
so/sov
oo/osv
1-2
2-1
11
11*
11
11*
16*
4
18*
2
Total
22
22
20
20
Response type “1-2” refers to the fKst noun being the agent, “2-1” where the second
noun is the agent. Correct responses are marked by an asterisk (*).
sequence as a sentence. Within each sentence form, children appear to overwhelmingly use the 1-2 response in OS/SOV, since the sequence exactly
mimics a simple SOV sentence. There was no tendency for children to prefer
the 2- 1 response in SO/OSV which mimics a simple OSV sentence, but this
is not surprising given that children were shown in the previous experiment
t o have difficulty in comprehending the OSV order. Thus, these errors indicate that children have a tendency to interpret the first noun as the agent,
and this is enhanced for the sequence which exactly mimics the canonical
SOV simple sentence.
Since Japanese children mis-segment the center-embedded sentences, these
sentences do not constitute an appropriate test of the effect of Focus and of
Word Order, since the relative clause is not properly parsed in these sentences. One can make specific tests of these effects in the left-branching sentences. The question of whether Subject Focus or Object Focus relative
clauses are easier t o comprehend can be answered by inspection of the number of correct responses in the subordinate clause for SS/SOV and OS/OSV,
.~
Relative clauses
Table 7.
2 13
Frequencies of responses on the main clause o f left-branching sentences
Response
Subject-Qbject-Verb
object-Subject-verb
SSlSOV
SOlSOV
os/osv
oo/osv
2-3
3-2
15*
6
17*
2
7
10*
8
11*
Total
21
19
17
19
“2-3” refers to responses where the second noun of the sentence is the agent, “3-2”
where the third noun is the agent. Asterisk (*) indicates correct response.
which are Subject Focus, and comparing this with SO/SOV and OO/OSV,
which are Object Focus. Response type “2-1” is correct for Subject Focus,
and “1-2” is correct for Object Focus. The Accessibility Hierarchy Hypothesis predicts Subject Focus to be better comprehended than Object Focus.
Based on our earlier experiments with comprehension of simple sentences,
however, an explanation based on the order of the subject and object predicts Object Focus to be easier, since it takes the SVO order. A breakdown
of the number of the response types with respect to the sentence forms appears in Table 6. As can be readily seen, there were more correct responses
on the Object Focus than on Subject Focus. This suggests that the order of
constituents is the major factor.
The effect of Word Order can best be assessed through comparison of
SS/SOV and SO/SOV with OS/OSV and OO/OSV on the number of correct
main clauses acted out, since these sentences are all left-branching. Table 7
displays the breakdown of the number of correct and reversed interpretations of the main clause for each sentence form. As can be seen, there were
more correct responses on the SOV order than on the OSV order, suggesting the operation of the SOV Constraint. The effects of Word Order and of
Focus will be tested explicitly in the next experiment.
A study by Harada et al. (1976) confirms the main findings of this experiment. The researchers tested act-out comprehension of SS, SO, OS and 00
in the SOV order in one group of children, and in OSV in a different group
of children. Thus, Word Order as a main effect was a between-subjects variable, while the other variables of Matrix Role and Focus were repeated measures. The researchers do not report statistical tests performed on their data,
but their subjects, ranging in age from 3 to 10, clearly found centerembedded sentences to be more difficult than left-branching sentences within
the SOV and within the OSV group. It is of interest to note that even at age
.
_
2 14
~
Kenji Hakuta
10, children have difficulty with centerembedded sentences, with less than
50% of the responses being correct.
Summary
This experiment tested children between 5;3 and 6;2 in their comprehension
of SS, SO, OS and 00 sentences in the SOV and OSV orders. The results
show an interaction between the Matrix Role of the complex noun phrase
and the Word Order of the sentence. Left-branching sentences were better
comprehended than center-embedded sentences. Thus, the grammatical description of the sentences did not account for the differential performance
of children on the various sentence types as a function of whether they were
in the SOV or OSV order. This result is a major embarassment to theories,
such as Parallel Function, that predict the psychological complexity of these
sentences in terms of their grammatical description. It appears that the configurational properties of the sentences constitute the variables to be considered in accounting for comprehension of these sentences. There was a
marginal effect for Word Order within the left-branching sentences, where
the SOV sentences were easier than the OSV sentences. In addition, relative
clauses whose head noun role was the object tended t o be easier than those
with subject head noun. However, even for the older age group, these sentences were inordinately difficult, and younger children were not tested.
Experiment two
The previous experiment demonstrated that 5 and 6 year old children find
left-branching sentences easier to comprehend than center-embedded sentences. Thus, S S and SO types were easier when they were in the SOV order,
but OS and 00 were easier when they were in the OSV order. However, even
for these children, the sentences proved quite difficult and the sentences required modification if younger children were to be tested.
The modification chosen for this experiment was to change one of the
clauses into an intransitive action, such as c a . When the intransitive action
is the subordinate clause, only the SS and OS types are possible, and since
the main clause is transitive, both these types can be either in the SOV or
OSV orders. Thus, for these sentences, it is posssible to test the effect of
Sentence Type (SS or OS) and the effect of Word Order (SOV or OSV). The
four sentence configurations are:
SSlSOV [CRIED] AGENT-ga PATIENT-o KICKED.
SS/OSV PATIENT-o [CRIED] AGENT-ga KICKED.
.
Relative clauses
2 15
OS/SOV AGENT-ga [CRIED] PATIENT-o KICKED.
OS/OSV [CRIED] PATIENT-o AGENT-ga KICKED.
As in the previous experiment, a main effect for Sentence Type would indicate the importance of the grammatical description, while an interaction between Sentence Type and Word Order would indicate the importance of
sentence configuration. A main effect for Word Order, with SOV better than
OSV, would indicate an extension of the SOV Constraint operating on simple
active sentences.
When the main clause is the intransitive action, it yields two sentence
types: SS and SO. The effect of Word Order and its interaction with Sentence Type cannot be tested with these sentences since word order change
can apply only to transitive actions, and the main verb in these sentences is
intransitive. However, with these sentences, we can test the effect of Focus,
since that parameter now varies. Thus, these sentences present a critical test
of the Accessibility Hypothesis. The two sentences are:
SS
SO
[PATIENT-o KICKED 1AGENT-ga CRIED.
[AG ENT-ga KICKED 1PA TIENT-ga CRIED.
If Keenan and Comrie’s accessibility hierarchy has any implications for processing difficulty of sentences in children, we would expect to find the effect
here, with better performance on SS th’an on SO. On the other hand, if comprehension of relative clauses were dependent on the linear order of elements,
then we would expect superior performance on SO over SS, since SO follows
the Subject-Verb-Object
order, while SS has the Object-Verb-Subject
order.
Subjects
Subjects were 36 children divided evenly into three age groups: Group 11:
3;3-4;2, Group 111: 4;3-5;2, and Group IV: 5;3-6;2.
Materials and procedure
The six structures described above were each replicated twice, creating a
total of 12 sentences for presentation to each child. The pool of words from
which the nouns and transitive verbs were selected was the same as in the
previous experiments. The intransitive verbs were: cried (naita), yawned
(akubi-shita), coughed (kushami-shita), and laughed (waratta). A second set
of sentences was created with the same intransitive verb but with the nouns
interchanged. The procedure was identical to that of previous experiments.
.
2 16
-
Kenji Hakuta
Scoring
A two-point scoring system was used, with one point given for correct performance on the transitive action and one on the intransitive action. Thus,
for each sentence, a score of 0, 1, or 2 was possible.
Results and discussion
The data were analyzed separately for the four structures in which the subordinate clause was intransitive and the two structures in which the main
clause was intransitive. The two analyses will be reported sequentially.
For the structures in which the subordinate clause was intransitive, a 3way ANOVA was performed, with Subjects nested within Age and the repeated measures factor, sentence, nested within a cross of Type (SS/OS)
by Order (SOV/OSV). Age did not turn out to be significant (F'[2,9] =
1.763), nor did it interact significantly with any of the repeated measures
factors. The only significant repeated measures factor was the interaction of
Type by Word Order ( F ' [1,2] = 22 1.457, p < 0.005). The means involved
in the 2-way interaction appear in Table 8. Inspection of the table reveals
that the interaction was due to good performance on SS/SOV and OS/OSV,
and poor performance on SS/OSV and OS/SOV. This result strongly s u p
ports the view that predicts sentence complexity on the basis of sentence
configuration. Japanese children frnd center-embedded sentences difficult
t o comprehend. That there was no main effect for Type (respective means
for SS and OS were 1.334 and 1.389) shows that the grammatical description of the sentences was not a good predictor of complexity. In addition,
the fact that there was no main effect for Word Order (respective means for
Table 8.
Means and standard deviations for Dpe by
Word Order interaction
Type
ss
os
Order
sov
osv
R = 1.639
s.d., = 0.407
s.d., = 0.039
d = 1.028
s.d., = 0.358
s.d., = 0.039
8 =1.069
8 =1.708
s.d., = 0.381
s.d., = 0.059
s.d., = 0.453
s.d., = 0.059
Relative clauses
Table 9 .
Breakdown of frequency of response types for SSlSOV, SSlOSV, OS/SOV
and OSlOSV
SS/SOV: [Vi] N-ga NOV
Transitive
SS/OSV: N-o [Vi] N-ga V
Intransitive
Total
Transitive
-
13*
10
36
10
49*
20
Total
23*
46
69
41*
13
6
0
53.
+
-
13
Total
60*
6
66
Transitive
OS/OSV: [Vi] N-o Nw V
Intransitive
+
Total
-
-
OS/SOV: N-w [Vi] N-o V
Intransitive
+
+
+
2 17
Total
Transitive
-
Intransitive
Total
+
-
+
12*
47
59*
+
55*
6
61*
-
0
4
4
-
6
1
I
12*
51
63
61*
7
68
Total
Total
The subordinate clauses are intransitive. "+" indicates correct response,
responses are marked by an asterisk (*).
"-"indicates
error. Correct
SOV and OSV were 1.354 and 1.368) suggests that the SOV Constraint may
not be operative for complex sentences. This will be discussed further after
an analysis of the errors.
A breakdown of the responses into how well the children comprehended
the transitive and intransitive components of the sentences is revealing. This
analysis excludes those responses for which only one clause was acted out,
although there were very few such cases. Responses can be classified into
four categories:
(1 )
(2)
(3)
(4)
Transitive
Transitive
Transitive
Transitive
correct, intransitive correct;
correct, intransitive error;
error, intransitive correct;
error, intransitive error.
Table 9 gives the frequency of the four types of response for each of the
structures. As can be seen from the column and row marginals, there was
very little difference across the structures with regard to the transitive clause,
21 8
Kenji Hakuta
the major difference being with respect to the intransitive action. The leftbranching sentences, SS/SOV and OS/OSV, showed good comprehension of
the intransitive clause, but there were more errors than correct responses for
the center-embedded structures SS/OSV and OS/SOV. The fact that the intransitive clause was centerembedded caused the children to erroneously
assign the first noun of the sentence as the subject of the intransitive verb.
This result is exactly what a left-to-right sentence processing model, such as
the Augmented Transition Network (e.g., Kaplan, 1975) would predict in
the form of “garden path” sentences (Wanner et al., 1974).
The fact that there is no apparent preference for SOV over OSV in the
comprehension of these sentences is not surprising once we realize that Word
Order for these sentences refers to the order of subject and object in the
matrix sentence. Consider the left-branching sentences with the configuration ViNNV, where Vi is the intransitive verb. As can be seen in Table 9,
most responses for SS/SOV and OS/OSV correctly assigned the first noun t o
the Vi. Thus, the first noun is indeed the subject, although not of the action
of the main clause. In addition, if we assume that the noun assigned to the
Vi is no longer the ‘‘first noun” for the processing of the remainder of the
sentence, there is no reason why SS/SOV should be easier than OS/OSV. In
fact, one would expect the OSV order to be easier since the second noun of
the entire sentence in effect becomes the “first noun” with respect to the
main verb. When the first noun is removed, the sequence SOV ends up as OV
while OSV ends up as SV. A different line of experiments explicitly testing
this possibility shows superior performance on OSV over SOV orders
(Hakuta, 1978), but that is beyond our current scope. The major point of
relevance for our present purpose is to show that in comprehending these
sentences, children appear to process the sentences quite locally. The SOV
Constraint may be applicable only when it is locally relevant, as in simple
sentences, at least for comprehension. In the case of the center-embedded
sentences, as Table 9 reveals, children almost unanimously chose the first
noun as the agent of Vi, erroneously, Thus once again, the first noun is the
subject of the subordinate action. The SOV Constraint is not readily apparent
in the context of their performance on the entire sentence because children
process the sentences locally.
Turning now to the SS and SO sentences in which the main clause was the
intransitive action, the results were analyzed by a 2-way ANOVA with Subjects nested within Age and crossed with the repeated measures factor, Sentence, nested within Type (SSlSO). Once again, Age was not significant
( F ’ [ 2 , 2 ] = 3.094), nor was its interaction with Type. There was a significant
main effect for Type ( F ‘ [1,5] = 9.568, p < 0.05), and-inspection of the
means revealed that children found the SO structure (X = 1.722, s.d., =
2 19
Relative clauses
Table 10. Breakdown of frequency of response types for SSand SO
SO: [N-ga VI N-ga Vj
S S : [N-o V] N-ga Vi
Transitive
Intransitive
Total
Transitive
+
Intransitive
+
-
Total
_
32*
29
6
3
38*
32
+
-
55*
5
13
3
68*
8
Total
61*
9
70
Total
60:
16
76
+
The main clause is intransitive and the subordinate clause is transitive. “+” indicates correct response,
indicates error. Correct responses are marked by an asterisk (*).
‘I-”
0.438, s.d., = 0.039) easier than SS ( g =1.403, s.d., = 0.428, s.d., = 0.098).
A breakdown of the responses into the four types as was done for the other
structures earlier, presented in Table 10, shows that the difference was mostly due to differential performance on the transitive action, namely the subordinate clause. This result does not support the hypothesis that Keenan and
Comrie’s accessibility hierarchy would have implications for children’s sentence processing. It is more consistent with the view that the local order of
elements within the complex noun phrase determines its ease of comprehension. The fact that subject focus is easier than object focus in English is best
explained by the fact that the head noun in English is located to the left of
the relative clause, and thus a subject focus sentence results in an S[VOl
configuration while an object focus sentence results in an O[SVl configuration. Since the head noun is on the right of the relative clause in Japanese,
subject focus results in an [OV] S order while object focus results in an
[SV] 0 configuration. We can derive a law which will predict ease of comprehension of subject and object focus in all languages:
In languages where the head noun is on the left of the relative clause, subject focus will be easier than object focus, whereas in languages where the
head noun is on the right of the relative clause, object focus will be easier,
all other things being equal.
Summary
This experiment simplified the sentences tested in the previous experiment
by making one of the clauses intransitive. This reduced the number of sentence types tested. SS and OS were tested in SOV and OSV orders with the
220
Kenji Hakuta
subordinate clause being intransitive, and S S and SO with the main clause intransitive. Subjects were between ages 3;3 and 6;2. A comparison of S S and
OS in both word orders strongly confirmed the interaction between Matrix
Role and Word Order found in the earlier experiment. Sentences were much
easier to comprehend when they were left-branching than when they were
centerembedded. In the centerembedded sentences, most children made the
error of interpreting the initial NVi ... sequence as a unit. There was n o main
effect for Word Order, and thus the SOV Constraint does not appear in the
context of the entire sentence. It was argued that children process the sentences at the local level. A comparison of S S and SO showed that Japanese
children find the relative clauses with subject focus more difficult than those
with object focus, contrary to the Accessibility Hierarchy Hypothesis. The
superior performance on object focus is attributable to an extension of the
SOV Constraint to the NVN sequence.
Experiment three
Up until this point, we have been using the terms “centerembedded” and
“left-branching” as convenient labels to capture the configurational properties of the sentences. This experiment will pursue the question of whether it
is centerembeddedness per se that causes difficulty in understanding a sentence, or whether the resultant configuration is the important variable. In
formulating the critical test, we will take advantage of the right-dislocated
structure frequently found in colloquial Japanese. In these sentences, a noun
phrase is placed to the right of the main verb of the sentence. Thus, a sentence with the canonical form AGENT-ga PATIENT-o BIT can be rightdislocated as AGENT-ga BIT, PATIENT-o or as PATIENT-o BIT, AGENTga. The main verb is marked in these structures by a terminal contour and,
in this experiment, it was decided to mark the verb with the dimunitive - E ,
which is frequently used in child speech and speech to children at the end
of verbs. Table 1 1 shows how right dislocation affects SS/SOV, SS/OSV, OS/
SOV and OS/OSV sentences. Although it is possible to dislocate the first or
the second noun phrase, in order to preserve the word order of the subject
and object, the second noun of each sentence was right-dislocated. To maintain consistency in notation, although the right-dislocated sentences are now
in the SVO and OVS orders, ‘I will refer to them as SOV and OSV. The original sentences in Table 1 1, under the left column labelled “Matrix NNV”, are
the common SOV and OSV forms. Thus, SS/SOV and OS/OSV are leftbranching while SS/OSV and OS/SOV are center-embedded and there is a
stacking of nouns. Under the right column labelled “Matrix NVN” appear
Relative clauses
221
Table 11. Illustration of how right dislocation affects sentence configuration
Type/Order
SSlSOV
ss/om
os/sov
os/osv
Matrix NNV
Matrix NVN
( N o V ]Ifga N-o V
IN-o V ]y-ga V, N-o
N-o V, IN-0 V ]y g
N - e V, [N-go V ]N-0
[N-go V]N-o V, N-ga
N-o IN-o VI c-ga V
N-ga [N-go V ]Pj-0 V
[N-go V ]N-o N - p V
Confwrations under “Matrix NNV” are the standard order. Configurations under
“Matrix NVN” are rightdislocated structures. Terminal contour at the end of the main
verb is indicated by a comma (,).
the right-dislocated sentences. Since Japanese places the relative clause to the
left of the head noun, the centerembedded sentences under Matrix NNV
still remain centerembedded under Matrix NVN. However, in these sentences, the nouns are unstacked such that there is no longer any local
mimicking of a simple sentence at the beginning of the sentences. In fact,
under Matrix NVN, the four sentences have the identical configuration,
NVNVN. Thus, the Matrix NVN sentences allow us to test whether centerembeddedness or configuration is the critical variable.
If centerembeddedness presents difficulty to the child, then there should
be a similar interaction between Sentence Type and Word Order in the
Matrix NVN sentences as there would be in the Matrix NNV sentences. On
the other hand, if the configuration of the sentence is important, there
should be no interaction between Type and Word Order in the Matrix NVN
senten ces.
Subjects
Subjects were eight children between ages 5;4 and 6 ; 3 .
Materials and procedure
Sentences with both the main and subordinate clauses reversible were used
for this experiment. Fortyeight sentences were constructed from a factorial
combination of two levels of Sentence Type (SSlOS), two levels of Word
Order (SOV/OSV), two levels of Matrix Form (Matrix NNV/Matrix NVN),
and four replications. In order to accommodate this large number of sentences, the pool of animals was increased. Sentences were grouped into four
blocks, with each block having one example of each sentence. In addition,
animals were also divided up according to these blocks. Presentation of
222
Kenji Hakuta
blocks as well as of sentences within each block was randomized for each
child. The animals in each block were introduced to the child immediately
prior t o the presentation of that block. This procedure helped maintain the
child’s interest through 48 sentences, and is highly recommended to other
researchers wishing to present large numbers of sentences to young children.
The entire procedure lasted about 30 minutes.
Scoring
Responses were scored 1 point for each clause correct, resulting in a total
possible score of 2 for each sentence.
Results and discussion
Overall analysis was conducted by a 4-way ANOVA, in which Subjects were
crossed with the repeated measures factor, Sentence, nested within a cross of
Type (SS/OS) by Word Order (SOV/OSV) by Matrix Form (NNV/NVN).
Main effects for Type and for Word Order were not significant (F’ < l ) , as
well as the main effect for Matrix Form ( F ’ [1,91 = 1.597, p < 0.25). The 2way interaction between Type and Word Order was significant ( F ’ [1,71 =
7.859, p < 0.05),as was the 3-way interaction of Type by Word Order by
Matrix Form ( F ’ [l,S] = 6.704, p < 0.05). The means, plotted in Figure 3,
Figure 3.
nree-way interaction between Matrix Order by Type by Word Order.
Matrix
NNV
: YI
Matrix N V N
c
1.41.2
OS
9
0
7
SOV
5s
osv
SOV
ORDER
OSV
Relative clauses
223
reveal that the 2-way interaction was due entirely to the interaction of Type
by Word Order within Matrix NNV, which also accounts for the 3-way interaction. In fact, when separate ANOVAs were performed by Matrix NNV and
Matrix NVN, the 2-way Type by Word Order interaction was significant for
Matrix NNV (F’[1,51 = 13.327, p < 0.05),and the mean square for Type
by Word Order in Matrix NVN was zero. It appears that the 3-way interaction was due to better comprehension of OS/SOV and SS/OSV, the centerembedded structures, in the Matrix NVN form, than in the Matrix NNV
form.
Since the predictions concern the variable of embeddedness (centerembedded or left-branching) and its relationship with the Matrix Form, the
two sentence configurations representing each level of embeddedness within
each Matrix Form were pooled. Thus, within the Matrix NNV structures, S S /
SOV and OS/OSV were pooled since they are both left-branching, and SS/
OSV and OS/SOV were pooled since they are both center-embedded. The
same pairs were pooled within the Matrix NVN. Pooling of these data is
justified on the grounds that it was planned a priori and, in addition, t’-tests
comparing each pair of means to be pooled did not produce any differences
that were even marginally significant.
The pooling of the data resulted in four means: LB(1eft-branching)/NNV
(2= 1.203, s.d., = 0.347, s.d.2 = 0.291); CE(centerembedded)/NNV (X=
0.750, s.d.1 = 0.189, s.d.2 = 0.259); LB/NVN (X = 1.125, s.d.1 = 0.320,
s.d., = 0.211); CE/NVN
= 1.125, s.d., = 0.284, ~ . d =. 0.259).
~
Comparison of these means through t‘-tests showed that although LB was easier than
CE within Matrix NNV (t’[4] = 4.231, p < 0.02), there was no difference
between LB and CE within the Matrix NVN ( t ‘ < 1). Thus, there is no difference between LB and CE when we control for sentence configuration. In
addition, the lack of difference between LB and CE within Matrix NVN is
due to CE sentences being easier in the Matrix NVN than in the Matrix NNV
form, as opposed to the other possibility, that LB might be more difficult
in Matrix NVN than in Matrix NNV. This is demonstrated by the fact that
while LB did not differ with respect to the two levels of Matrix Form, CE
sentences were significantly easier in Matrix NVN than in Matrix NNV ( t ‘[ 5 I
= 3.487, p < 0.02).
Individual analysis of the responses showed that seven out of the eight
children had better comprehension of CE when it was in Matrix NVN than
Matrix NNV. Performance was better on Matrix NVN because of an increased
comprehension of the subordinate clause, with a concommitant decrease in
the garden-path type errors involving the first two nouns of CE/Matrix NNV
sentences. The reader will recall that children erroneously segment the initial
NNV ... sequence as a sentence in these structures. An “improvement score”
(x
224
Kenji Hakuta
was calculated for the subordinate and main clauses for the two CE sentences, which consisted of the number of correct responses for a given clause
in Matrix NVN minus the corresponding number in Matrix NNV. For SS/
OSV, the improvement score for the subordinate clause was +12, while for
the main clause it was -1. There was a reduction by 1 1 of the garden-path
type errors. For OS/SOV, the improvement score for the subordinate clause
was +11, while for the main clause it was +2. The reduction in garden-path
type errors was 12.
The results of this experiment show that the difficulty of center-embedded
sentences in the Matrix NNV form can be removed through right dislocation,
which prevents the stacking of nouns that results in garden-path errors.
Centerembeddedness per se is not the critical variable in comprehension in
children. The center-embedded sentences in the Matrix NVN form were no
more difficult t o comprehend than left-branching sentences. I claim that
centerembedded sentences tend to be difficult because they frequently
cause stacking qf nouns and verbs which need to be re-ordered before the
sentence can be interpreted. Such is the case in the English SO form, such as
The donkey that theman kicked licked the mule. Center-embedded sentences
are not difficult when the nouns are stacked such that they are separated out
by verbs, as in the English S S : The donkey that licked the man kicked the
mule. In Japanese as well, centerembedded sentences are difficult when the
stacking causes an erroneous assignment of nouns to verbs, as in the sentences with the NNVNV configuration. They are no more difficult than leftbranching sentences, however, when the stacking is “neatly” done, as in
NVNVN. It is possible that the obligatory terminal contour at the end of the
main verb in rightdislocated sentences will cause them to be perceived as
two left-branching sentences rather than as a center-embedded structure.
Nevertheless, since the transitive action of the main verb involves both the
noun to its left and the complex noun phrase to its right, they are still best
considered as center-embedded.
Summary
This experiment tested the hypothesis that center-embeddedness increases
difficulty of sentences, which we have assumed so far, against the proposal
that the stacking o f nouns in such a way that they are interpreted erroneously would cause difficulty in comprehension. The second matrix noun of
centerembedded sentences OS/SOV and SS/OSV and left-branching sentences SS/SOV and OS/OSV were right-dislocated, yielding a uniform unstacked configuration for both the left-branching and center-embedded sentences. Children between ages 5;4 and 6;3 were tested on comprehension of
Relative clauses
225
reversible sentences created in this manner. The results show that centerembedded sentences are no more difficult than left-branching sentences
when the nouns are unstacked. Thus, it is not centerembeddedness per se
that causes processing difficulty, but rather its frequently concommitant
stacking of constituents.
Experiment four
The results from the studies in comprehension of complex sentences showed
that Japanese children have difficulty with sentence forms that take the configuration NNVNV. For this finding to be of any generality, it is important
to ask the question as to whether similar difficulties are manifested with the
configuration in other linguistic tasks. The present study looks at children’s
immediate imitation and delayed production of these sentences. In immediate imitation, the prediction is that the NNVNV configuration should be
more difficult to imitate than the NVNNV configuration, since the former
is stacked and the latter is unstacked. In delayed production, the children
will produce sentences in the NVNNV configuration and avoid the NNVNV
configuration.
The immediate imitation and delayed production data were obtained in
the same task. Sentences in the immediate imitation phase were paired with
pictures depicting the action. The same pictures were used to cue the delayed production of the sentences. The pictures were constructed such that
they depicted actions described by SS and OS sentence types. In the SSpicture, the agent of a transitive action is shown performing another action,
while in the OS-picture, the patient of a transitive action is shown performing another action.
In the immediate imitation phase, the model sentences came in either the
SOV or OSV order for each of the pictured types. Thus, there were four
model sentence forms: SS/SOV, SS/OSV, OS/SOV, and OS/OSV. The assumptions underlying the immediate imitation task are that sentences which
are easier for the child will be more correctly imitated, and that errors will
reflect changes towards their preferred form of the sentences.
In the delayed production phase of the task, within the rough constraint
of producing sentences with relative clauses, the child should change the
order of the sentences to the preferred form.
The prediction is that if children find the NNVNV Configuration difficult
t o produce, given that they have the choice of producing any given sentence
in the SOV or OSV order, for the two pictured meanings SS and OS, they
should choose the order which takes the configuration NVNNV, which is
226
Kenji Hakuta
unstacked. Thus, for the SS-picture, there should be a tendency to produce
the SOV order, while for the OS-picture, the OSV order should be preferred.
However, since SOV is dominant in Japanese, there may be a tendency for
children to produce stacked structures for OS, since it preserves the basic
word order. Although word order did not appear as a main effect in the comprehension results, I argued that this was due to the fact that children processed the sentences locally. For this task, to the extent that children process
the matrix sentence, one might expect an effect for word order in imitation
and in delayed production.
Subjects
Subjects were 36 children divided into three age groups: Group 11:
4;8, Group 111: 4;9-5;8, and Group IV: 5;9-6;8.
39-
Materials and procedure
The pictures used as cues for the sentences contained three animals, of which
two were identical. Two identical animals were pictured so that there was a
natural reason for using relative clauses, highlighting their restrictive function.
The third animal will be referred to as the unique animal. One of the identical
animals was pictured performing either an intransitive (crying or yawning) or
a transitive (eating an apple or holding a balloon) action. This action corresponds to the relative clause. For the SSpictures, this animal was pictured as
the agent of a transitive action (hitting, kicking or pushing) with the unique
animal. For the OS-picture, the animal was the patient of the transitive
action. The SS-pictures and the OS-pictures were balanced to evenly represent the relative clause actions. Two replications of each of the resulting
eight combinations were created, using different animals. This produced a
total of 16 pictures. An additional counterbalancing procedure involved an
equal placement of the transitive action on the right and left part of the
picture. The pictures were individually mounted on 8” X 11” construction
paper.
From each picture type, half of the pictures were chosen to be paired with
a model sentence in the SOV order, and the other half in the OSV order.
Within each order, half contained a relative clause with the intransitive verb
and the other half with the transitive verb. A second set of sentences was
created reversing the order of the sentences paired with each picture. Children were assigned randomly to either set. For each child, order of presentation of the picture/sentence pairs was randomized.
Each child was introduced to the pictures, which were bound together
with ring binders and made to resemble a story book, and informed that
-
~
.
- -
-__
-
.
..
.
.
..
Relative clauses
227
s/he was to help tell stories to a puppet. The child was told to repeat the
story just like the experimenter, always my mother, tells it. There were
four warm-up pictures of simple actions. If s/he did not repeat the sentences
verbatim, the child was instructed through modelling by the two experimenters, an example of mother and child interaction. Children encountered
no difficulty with the task, with the exception of one child who insisted on
telling her own story about the pictures. She was not included in the experiment. The immediate imitation phase of the experiment immediately followed. This part of the experiment lasted about 1 5 minutes. Following completion of this phase, the child was told that it was now his/her turn to tell the
story, and that the experimenter would not say anything. Children in general
readily took to this task. The pictures were shown in the same order as in the
immediate imitation phase. The delayed production phase also lasted about
' 1 5 minutes. No child failed to complete the task, and in fact many wanted
to d o more. Thus, the task was intrinsically interesting to the children. The
entire session was recorded on a cassette tape recorder and subsequently
transcribed.
Scoring
The response on the immediate imitation phase was coded as correct (1) or
incorrect(O), and in addition, errors were coded if the wrong particle was
used, and if there was a transposition of word order.
The delayed production data were recorded as being in the SOV or OSV
order if they contained a relative clause. If there was an error in the use of
particles, this was separately noted. In addition, a separate category was
created for those responses in which the child produced a conjoined sentence
involving the two actions.
Results and discussion
Since separate analyses were conducted for the immediate imitation and the
delayed production phases of this experiment, the results will be reported
separately.
Immediate imitation
Overall analysis for the immediate imitation data was conducted by a 4-way
ANOVA, with Subjects nested within Age, crossed with repeated measures
factor, Sentence, nested within a cross of Type (SS/OS) by Word Order
(SOV/OSV) by Verb Type (Transitive/Intransitive). Age was significant as a
228
Kenji Hakuta
Table 12. Means and standard deviations of means in 5 p e by
Word Order interaction in immediate imitation task
sov
osv
ss
a
f
os
f
=0.944
s.d., =0.135
s.d., = 0.051
= 0.660
s.d., = 0.317
s.d., = 0.105
=0.375
s.d., = 0.330
s.d., = 0.086
f =0.660
s.d., =0.317
s.d., = 0.070
main effect (F'[2,26] = 8.712, p < O.Ol), indicating overall better performance in older children.
The means for the four sentence forms summing across the age groups are
shown in Table 12. In the overall ANOVA, Type was not significant (F' < 1),
but there were significant effects for Word Order ( F ' [1,101 = 68.269, p <
0.001) and for the Type by Word Order interaction ( F ' [1,101 = 10.145, p <
0.01). Since there was no difference between the means for OS/SOV and
OS/OSV, these significant effects can be completely accounted for by the
large difference between SS/SOV and SS/OSV. In more ordinary language,
this means that while there was no difference whether an OS sentence was in
the SOV or OSV order, it made a large difference for SS, which was much
easier when it was SOV than when it was OSV. This is not totally in accord
with the comprehension results, which would predict that OS in the OSV
order should be easier since it is unstacked. But the result is not surprising
when one considers the possibility that there could have been a trade-off between the Stacking Constraint and the SOV Constraint, since the unstacked
configuration for OS is in the OSV order. This account of the data is supported by the errors made by the children in imitation, and will be discussed
shortly. It appears that in this imitation task, perhaps because of the cue to
meaning provided by the pictures, children processed the main verb t o the
extent that the effects of the SOV constraint were observed.
There were significant interactions of the repeated measures factors with
Age that were of interest. Age interacted with Type (Age X Type, F'[2,11]
= 4.1 16, p < 0.05) and with Type by Word Order (Age X Type X Word
Order, F'[2,14] = 4.901, p < 0.05). However, these interactions can be explained by the fact that even the youngest children performed close to optimum on the SS/SOV sentences, while there was improvement with age on
all the other sentence types.
~-
- -
Relative clauses
229
Table 13. Frequency of (a) particle errors and (b) order
transposition errors in immediate imitation task
(a)Particle errors
Type
Order
Total
sov
osv
ss
os
3
6
41
38
44
44
Total
9
79
88
(b) Order transposition errors
Type
Total
Order
sov
osv
ss
os
0
11
24
0
24
11
Total
11
24
35
The remaining main effect, for Verb, also turned out significant ( F ’ [1,31
< 0.02), with better performance on the sentences with intransitive verbs then transitive verbs. This is easily accounted for by the fact that
the sentences with intransitive verbs were shorter. In fact, the interaction of
Age by Verb is significant at the alpha level of 0.10 ( F ’ [ 2 , 6 ]~ 4 . 1 3 8 and
)~
inspection of the means reveals that while sentences with the intransitive
verbs are easier for Groups I1 and 111, the difference disappears by Group IV.
An analysis of the distribution of error types with respect t o the four
sentence forms is highly revealing as to where the children encountered difficulty in imitation. The number of errors on each sentence form appear in
Table 13. Table 13(a) gives the frequency of particle errors. For both sentence types, the particle errors appear almost entirely in the OSV order. This
consisted of changing the particle on the first noun, -e,to -ga. This error reflects the general preference, found also in simple sentences, for the SOV
order. Table 13(b) gives the frequency of transposition errors, where the
order of the constituents was switched. Transposition errors occurred exclusively for sentences in which constituents were stacked, namely SS/OSV
and OS/SOV. This error reflects the tendency t o create unstacked NVNNV
sequences out of stacked ones.
= 22.795, p
230
Kenji Hakuta
Delayed production
The delayed production data were analyzed in three separate ANOVAs, each
using a different measure since there is no “correct” performance on this
phase of the experiment. ANOVA 1 used the SOV order as the measure.
Each sentence in the SOV order was given a value of 1, with all other forms
assigned a value of 0. Subjects were nested within Age, crossed with repeated
measures factor Sentence nested in a cross of Type (SS/OS) by Model (SOV/
OSV), where model referred t o the original word order in which the sentence
was presented in the immediate imitation phase of this experiment. Model
was not expected to have a significant effect. ANOVA 2 used the OSV order
as its measure. Each sentence in the OSV order was assigned a value of 1, and
all others 0. The structure of ANOVA 2 was identical to ANOVA 1. In
ANOVA 3, the measure was sentences in the stacked configuration, where a
value of 1 was assigned to each sentence in the NNVNV configuration, namely, OSV for SS and SOV for OS, and all other forms were assigned a value of
0. Since the effect of Model was already tested in ANOVAs 1 and 2, this
ANOVA consisted of Subjects within Age crossed with Sentence within
Type (%/OS).
The predictions were that there should be significant effects for Type in
ANOVAs 1 and 2, but in opposite directions. In ANOVA 1, since the measure was SOV, a higher score was expected for SS than OS. For ANOVA 2,
where the measure was OSV, a higher score for OS was expected. These predictions were based on the assumption that stacked sentences will tend to be
avoided in delayed production. ANOVA 3 tested the prediction that when
stacked configurations are produced, they tend to be produced in the SOV
order, reflecting the trade-off between the preference for the SOV order and
avoidance of stacked configurations. Thus a main effect for Type was expected, with a higher score for SS since it is in the SOV order, while the OS
is in the OSV order. And finally, while an overall main effect for Age was
expected, this effect was considered uninteresting since it was also expected
that older children would tend to produce more relative clauses in their production overall, and the main effect would simply reflect this fact. While an
Age by Type interaction was not predicted, this would be a theoretically
more interesting effect open t o interpretations.
The results of ANOVA 1 showed significant main effects for Age (F’
12,271 = 6.612, p < 0.005) and for Type (F‘[1,291 = 72.119, p < 0.001).
The direction of the effect for Type was in the predicted direction, with a
higher score on SS
= 0.764, s.d., = 0.286, s.d., = 0.053) than on OS
=
0.205, s.d., = 0.262, s.d., = 0.066). The Age by Type interaction was not
significant (F’ < 1). The results of ANOVA 2 showed Type once again signi-
(x
(x
’
-
Relative clauses
231
ficant as a main effect (F‘[1,261 = 91.704, p < 0.001), and Age was approaching significance (F’[2,201 = 3.026, p < 0.10). The direction of the
effect for Type was again in the predicted direction, with a higher score on
OS
= 0.622, s.d., = 0.284, s.d., = 0.070) than on SS
= 0.056, s.d., =
0.132, s.d., = 0.039). The Age by Type interaction was not significant (F’<
1). An unexpected result consistent across both ANOVAs was the marginally
significant effect for Model. For ANOVA 1, F’(1,12) = 3 . 5 3 3 , ~< 0.10, and
for ANOVA 2, F’(1,lo) = 3.392, p < 0.10. In both cases, the effects were in
the direction predicted if children retained the word order of the sentence as
it was originally presented. Inspection of individual sentences and subject responses did not reveal any outliers or biases on particular items, and thus it
must be concluded that there was a marginal retention of sentence form.
This does not influence the interpretation of the other results, however,
since Model did not interact significantly with any of the other factors in
either of the ANOVAs. The results strongly indicate that children overwhelmingly prefer to produce sentences that are unstacked.
The results of ANOVA 3, however, indicate that there is also a significant
trend to produce stacked sentences in the SOV order. This is indicated by a
main effect for Type (F‘[1,27] = 13.957, p < 0.001). Thus, there was a
higher value for OS/SOV
= 0.205, s.d., = 0.262, s.d., = 0.066) than for
SS/OSV
= 0.056, s.d., = 0.132, s.d., = 0.039). The effects for Age (F‘
12,341 = 1.327, p > 0.10) and the Age by Type interaction (F‘[1,27] =
1.365, p > 0.10) were not significant.
For the reader’s convenience, the relative frequencies of the four sentence
forms, SS/SOV, SS/OSV, OS/SOV and OS/OSV, are displayed in Table 14,
where a score of 1 is given to each occurrence of the sentence form. The
results from the delayed production phase of this experiment show that
there are two constraints operating on the child in the production of sentences containing relative clauses. The first constraint is that they will tend
to produce sentences with unstacked configurations rather than stacked configurations, the Stacking Constraint. The second constraint, the SOV Constraint, is that they will tend to produce sentences in the SOV order. Thus,
stacked sentences are produced for the reason that they fulfill the tugging
demands of this latter constraint. This conclusion is supported by inspection
of the data from individual children. Out of the 36 children in this experiment, 21 produced at least one stacked OS sentence (which is in the SOV
order), with 6 of them producing four or more. But only 8 children produced any stacked SS sentences.
An inspection of the distribution of errors involving particles reveals a
similarity to particle errors in the immediate imitation phase of this experiment. Table 15 gives a breakdown of the error rates for each of the four
(x
(x
(x
(x
.
232
Kenji Hakuta
Table 14. Relative frequencies o f the four sentence forms
produced in the delayed production task
sov
osv
~
ss
os
~~
f
= 0.164
f
s.d., = 0.286
s.d., = 0.053
=0.056
s.d., ~ 0 . 1 3 2
s.d., = 0.039
f
2
= 0.205
s.d., = 0.262
s.d., = 0.066
=0.622
s.d., = 0.284
s.d., = 0.070
Numbers represent average per response, where a score of 1
is given for each instance of a given form produced.
Table 15. Percentage of particle errors in sentence forms
produced in the delayed production task
ss
os
sov
osv
1%
(161220)
50%
(8116)
2%
(1159)
25%
(4411 79)
sentence forms. As can be readily seen, particle errors are clustered around
the OSV order. Most of these errors, as in the immediate imitation phase, involved marking the first noun of the sentence with -ga. This is yet another
manifestation of the toll taken by the SOV Constraint when the sentences
were produced in the OSV order, in particular for the OS/OSV, since its
appearance in the OSV order is the result of the Stacking Constraint.
Coordination provides an alternative way of describing the pictured events.
Fourteen children produced 58 coordinated sentences in all, of which 39
were produced by the youngest age group (Group II), and 6 and 13 by
Groups 111 and IV respectively. Thus, younger children tended to produce
coordinate structures. There were no strong trends towards producing more
coordinations with respect to either picture type. There were 32 coordinations for OS-pictures, as compared to 26 for SS-pictures. One consistent
trend was for the children to mention the action corresponding to the sub-
-
~
~
__
-
Relative clauses
233
ordinate clause first. This is not terribly surprising since the subordinate
clause actions were quite distinct from the main clause actions, and the subordinate clause always preceded the main clause verb in the modelled sent ences .
In order to ensure that the results of this study were not a function of the
pictured meanings per se, but rather resulted from the processing constraints
inherent in the sentences required for their description, a control group of
14 children between 3;8 and 6;8 was set up. The procedure was identical to
the experiment just described, except that simple sentences were paired with
the pictures, without making reference to the subordinate clause action.
In this control experiment, children overwhelmingly preferred the SOV
order for simple sentences in both the immediate and delayed imitation phase
of the task. Most importantly for our present purposes, the effect of the
picture Type was tested in order to determine whether the pictures meaning
might have influenced the results obtained in the main experiment. Neither
the main effect for Type, nor its interaction with Model, was significant
(F' < 1). Thus, the results obtained using complex sentences cannot be attributed to the possibility that Japanese children generally prefer the SOV
order for SS-pictures and OSV order for OS-pictures.
Summary
This experiment used an immediate imitation and delayed production paradigm to test whether children showed the SOV Constraint and the Stacking
Constraint in their production. Subjects were between 3;9 and 6;8. Children
found SS sentences much easier to imitate correctly when they were in the
SOV order than when they were in the OSV order, while no strong preference for word order was found for the OS sentences. Overall, both the Stacking Constraint and the SOV Constraint contributed to the results obtained.
Errors in imitation strongly supported this conclusion. The delayed production results also pointed to the contribution of both constraints. Children
had a strong preference for producing both SS and OS in their unstacked
word order, but the tendency t o produce stacked sentences was stronger for
OS than for SS because the former is in the SOV order. A control condition
showed that the results obtained were not a function of the pictured meanings inherent in SS and OS.
Summary and conclusion
The confused state of the art in the investigation of the comprehension of
complex sentences containing relative clauses in English was a primary moti-
234
Kenji Hakuta
vation underlying the investigation of complex sentences in Japanese children. The hypotheses that had been advanced to account for the English
data were classified into two classes depending on their underlying assump
tions. One class assumes that the grammatical description of the sentences is
the appropriate basis from which to predict sentence complexity for children. The other class of hypotheses assumes the configurational properties
of the sentences to be the important predictors of sentence complexity. It is
impossible to unconfound these two classes of explanations in English for
the simple reason that it is a language where the grammatical description determines the sentence configuration, since it has a rigid word order. It is
therefore necessary to look at other languages. Japanese is an ideal language
because of its variable word order, thus making it possible to unconfound
the relative contributions of grammatical description and sentence configuration towards predicting sentence complexity. The data from Japanese
children unequivocally point to sentence configuration as the appropriate
variable. Whenever stacking of constituents occurs in a sentence such that
there is local mimicking of simple sentences, children appear to find the
simple sentence interpretation irresistible. When such stacking does not
occur, even center-embedded sentences are not particularly difficult. These
results allow us to reject the Parallel Function Hypothesis and the Accessibility Hierarchy Hypothesis for any language unless they are re-stated in such a
way that they interact with the configurational demands of particular languages. In fact, any hypothesis stated purely in terms of the grammatical description of sentences is challenged by these results.
The question arises as to the generality of these findings in Japanese to
other SOV languages as well as to VSO languages. The results reported in this
paper suggest a somewhat secondary role for particles in the processing of
complex sentences, with a greater reliance on the configurational patterns of
constituents. There is, however, another variable that may need to be taken
into account in looking at differences within SOV and within VSO languages.
Within these two rough categories of languages, there are different patterns
of correlations between particles and their expected position within sentences. For example, in the case of Japanese simple sentences, there is a predominant SOV order such that the subject-marking particle -ga most
frequently appears on the first noun of the sentence. This correlational property has been observed to have a significant effect on the comprehension of
simple sentences in Japanese children (Hakuta, 1977). Slobin (1 978) reports
data on similar simple sentences from Serbo-Croatian and Turkish, both SOV
languages but with different patterns of correlation between inflections and
word order. Serbo-Croatian, like Japanese, is apparently a language where
word order is relatively free but not quite as free as Turkish in which inflec-
- - ___
\
- .-- - - - -
-
-.
-
Relative clauses
235
tions are highly regular and obligatory. Slobin reports that the Turkish children had little difficulty with varying word orders, relying primarily on particles for comprehending sentences. The Serbo-Croatian children, on the
other hand, reportedly “require normal marking in terms of both word order
and inflection for comprehension” (p. 24). It is thus possible that the
pattern of evidence for relative clauses within SOV languages may show interesting variation depending on the extent t o which the inflectional system
and word order interact. Future reports on relative clauses from Slobin’s project on Serbo-Croatian and Turkish promise to shed light on the nature of
this variation.
The fact that stacked sentences that would ordinarily cause perceptual difficulty can be simplified through word order variation suggests the possibility
that flexible word order might itself be a device invented by languages that
are plagued by such structures. Antinucci et al. (1979) suggest that SOV languages have such difficulty, which may cause the diachronic change from
SOV to SVO. It may be the case, however, that SOV languages might as a
consequence evolve flexible word order, which would either delay or obviate
the need for change into SVO. Independent linguistic analysis should reveal
the distribution of the extent of word order variation allowed by languages
across different basic underlying structures.
References
Antinucci, F., Duranti, A. and Gebert, L. (1979) Relative clause structure, relative clause perception,
and the change from SOV to SVO. Cog, 7,145-176.
Bever, T. G. (1970) The cognitive basis for linguistic structures. In J. R. Hayes (ed.),Cognition and the
Development of Languuge. New York, Wiley.
Brown, H. D. (1971) Children’s comprehension of relativized English sentences. Child Devel., 42,
1923-1936.
Chomsky, C. (1969) The Acquisition of Synrux in Qlildren from Five t o Ten. Cambridge, MA, MIT
Press.
Chomsky, N (1961) On the notion “rule of grammar”. In Proceedings of the Twelfth Symposium in
Applied Muthematics, American Mathematical Society, pp. 6-24.
Clark, H. H. (1973) The language-as-fixed effect fallacy: a critique of language statistics in psychological research. J. verb. Learn. verb. Behav., 12, 335-359.
de Villiers, J. G., Tager-Flusberg, H., Hakuta, K. and Cohen, M. (1979) Children’s comprehension of
relative clauses. J. psycholing. Res.
Greenberg, J. H. (1963) Some universals of grammar with particular reference to the order of meaningful elements. In J. H. Greenberg (ed.), Universals of Longuuge. Cambridge, MA, MIT Press.
Hakuta, K. (1977) Word order and particles in the acquisition of Japanese. Pupers and Reports on
Child Longuuge Development, 13,110-117.
Hakuta, K. (1978) Constraints on children’s comprehension of complex sentences. Paper presented at
the Workshop Conference on Language Acquisition, University of Massachusetts, Amherst,
April 1978.
.~
236
-
-
Kenji Hakuta
Harada, S. I., Uyeno, T., Hayashibe, H. and Yamada, H. (1976) On the development of perceptual
strategies in children: a case study on the Japanese child’s comprehension of the relative clause
constructions. Annual Bulletin of the Research Institute Logopedics Phoniatrics, University of
Tokyo, 10, 199-224.
Kaplan, R. (1975) Transient Processing Load in Sentence Comprehension. Unpublished doctoral dissertation, Harvard University.
Keenan, E. L. (1975) Variation in universal grammar. In R. W. Fasold and R. W. Shuy (eds.), Analyzing Variation in Language. Washington, DC, Georgetown University Press.
Keenan, E. L. and Comrie, B.(1972) Noun phrase accessibility and universal grammar. Paper presented
at the Linguistic Society of America, Winter.
Kuno, S. (1973) The Structure of the Japanese Language. Cambridge, MA, MIT Press.
Kuno, S. (1974) The position of relative clausesand conjunction. Ling. Inq., V, 117-136.
Lehmann, W. P. (1973) A structural principle of language and its implications. Lang., 49. 47-66.
McCawley, N. (1976) Reflexivization: a transformational approach. In M. Shibatani (ed.), Syntax and
Semantics: Japanese Generative Grammar, Vol5, New York, Academic Press.
McNeill, D. (1966) The creation of language by children. In J. Lyons and R. J. Wales (eds.), A y c h o linguistic Papers. Edinburgh, Edinburgh University Press.
Miller, G. A. (1962) Some psychological studies of grammar. A m . Psychol., 17, 748-762.
Ross, J. R. (1967) Constraints on Variables in Syntax. Unpublished doctoral dissertation, MIT.
Ross, J. R. (1973) The center. Paper presented at New Ways of Analyzing Variation in English
( W A V E ) Conference, Georgetown University, Washington, DC.
Schachter, P. (1976) The subject in Philippine languages: topic, actor, actor-topic, or none of the
above. In C. Li (ed.), Subject and Topic, New York, Academic Press.
Sheldon, A. (1974) The role of parallel function in the acquisition of relative clauses in English. J.
verb. Learn. verb. Behav., 13, 272-281.
Slobin, D. I. (1973) Cognitive prerequisites for the development of grammar. In C. A. Ferguson and
D. I. Slobin (eds.), Studies of Child Longuage Development. New York, Holt, Rinehart and
Winston.
Slobin, D. I. (1978) Universal and particular in the acquisition of language. Paper presented at workshopconference on Language Acquisition: State of the Art, University of Pennsylvania, May,
19-22.
Smith, M. (1974) Relative clause formation between 29-36 months: a preliminary report. Papersand
Reports on sild Language Development, 8, 104-110.
Tavakoliin, S. (1978) The conjoinedclause analysis of relative clauses and other structures. In H.
Goodluck and L. Solan (eds.), Papers in the Structure and Development of Child Language,
University of Massachusetts Occasional Papers in Linguistics, Vol. 4,1978.
Wanner, F.,Kaplan, R. and Shiner, S. (1974) Garden paths in relative clauses. Unpublished manuscript,
Harvard University.
Winer, B. J. (1971) Statistical Principles in Experimental Design, 2nd ed., New York, McGraw-Hill.
Risumi
Le choix d’une description de la complexit6 des phrases en termes de relations grammaticales ou en
termes des propri6t6s configurationnelles de la structure de surface est un point fondamental des
psycholinguistiques du d6veloppement. Ce problkme a fait I’objet de recherches en anglais mais les
&tudes, 6tant donne les propriitks particulieres de I’ordre des mots dans cette langue, n’ont pas 6clair6
la question. Une skrie d’expiriences men6es avec des enfants japonais a permis de montrer que les exigences confgurationnelles de la phrase reprisentent le facteur critique.